Semantic similarity-based validation of human protein-protein interactions

In this study, we perform a quantitative assessment on the application of gene ontology-based semantic similarity measures in human protein interaction analysis. The ability of different similarity measures to discriminate true positive and false positive protein interactions is assessed using receiver operating characteristic (ROC) graphs. Our results indicate that two semantic similarity measures based on GO biological process and molecular function annotation can be used to stratify human protein interactions. However, the similarity measurement based on GO cellular component annotation may not have the ability to discriminate true protein interactions from false positives. The combination of two measures yields better performance as compared to using either GO category alone for the classification. In addition, GO-derived semantic similarity measures have been shown to be valuable for the characterization of biological pathways. We believe that the integration of these measures with other information such as gene expression and network topology would greatly reduce the error rate in high-throughput protein interaction data.